Research On Superpixel Segmentation Algorithm Based On Ball K-means Clustering

Image segmentation,a pre-segmentation technique in computer vision processing,aims to divide an image into several disjoint sub-regions with high similarity.While pixel-level segmentation methods are no longer applicable to image processing,superpixel segmentation methods not only preserve the local information of the image itself,but also improve the speed of subsequent processing,and have been widely used in many fields.Since most of the existing superpixel segmentation algorithms still need to improve the edge fit and reduce the complexity,this paper investigates the Ball k-means clustering algorithm and proposes an iterative superpixel segmentation algorithm based on the Ball k-means clustering.To further investigate the applicability of superpixel segmentation results,this superpixel segmentation method is applied to salient object detection,the specific results are as follows:(1)Iterative algorithm for superpixel segmentation based on the Ball k-means clusteringFor the superpixel segmentation problem,in order to further improve the superpixel edge fit,the Ball k-means clustering algorithm is used for image segmentation,and an iterative algorithm for superpixel segmentation based on the Ball k-means clustering is proposed.First,the algorithm treats each superpixel as a five-dimensional hypersphere,and calculates the center and radius of the initial superpixel obtained by uniform segmentation;second,the nearest neighbor relationship among superpixels is obtained based on the relationship between the distance of the centers and the radius,and this is used as the partitioning criterion;finally,the superpixels are partitioned and clustered,and so on iteratively to achieve superpixel segmentation.Experiments show that,on the premise of the same number of superpixels,compared with other comparison algorithms on BSD500 data set,the edge fitting rate of the proposed algorithm is improved by 10% on average,and the segmentation accuracy is improved by 0.9%.Therefore,the algorithm improves the superpixel segmentation effect and stability,and obtains segmentation results with better fit to the real boundary of the object and clearer contours.(2)The salient object detection algorithm based on the Ball k-means clusteringTo further highlight the salient object and improve the integrity of detection,superpixels are applied to salient object detection,and a salient object detection algorithm based on the Ball k-means clustering is proposed.The algorithm is divided into two stages,the first stage adopts the idea of the Ball k-means clustering to generate superpixels;the second stage is to extract the saliency map.Firstly,the background salient value is obtained based on the color difference and texture features of the image boundary superpixels from a local point of view;secondly,the foreground salient value is obtained using the color difference of all superpixels;finally,the spatial variance is used to fuse two kinds of salient values,and the final salient map is obtained by normalization.Experiments show that the algorithm relies on superpixels based on the Ball k-means clustering for detection,and the detection integrity and prominence effect of the salient object region are better.The background suppression is more thorough,and the algorithm is insensitive to parameters and more reliable in detection.